PO.BCS02.03 · 生物信息与计算
多尺度基础AI描述符实现数字化肾细胞癌病理中肿瘤的精确定位
Multi-scale foundational AI descriptors enable accurate tumor localization in digitized renal cell carcinoma pathology
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：肾细胞癌（RCC）的诊断依赖于在病理全切片图像上对肿瘤的精确定位，但存在显著的观察者间变异性，可能影响治疗选择。数字病理学AI工具的最新进展催生了基础模型（foundation models，FMs），这些模型在多样化的H&E图像上训练以学习组织形态学模式。然而，这些模型的性能仅在单一尺度上进行评估，而临床病理学评估则利用多种尺度和放大倍率。我们的目的是开发并验证用于肾癌肿瘤定位的多尺度病理FM特征。
方法：从公共数据库中整理了RCC的H&E染色全切片图像。对于每个病例，在三种组织学尺度（5×、10×和20×放大倍率）上采样组织区域。使用病理FM MUSK，从每个区域及每个尺度生成组织形态学的定量表征，并使用在现有病理学家注释上训练的监督分类器区分肿瘤与非肿瘤组织。通过整合每个区域三种放大倍率的特征表征，构建了多尺度特征。我们使用其他FMs（CONCH、HOPTIMUS）并与在单个放大倍率上训练的模型进行比较，从而验证了我们的发现。
结果：共纳入118张切片（2家机构），按患者划分为训练集（n = 94）和留出验证集（n = 24）。切片经采样产生了300,000张以上标注图块（约112,000张肿瘤和约222,000张非肿瘤）。多尺度MUSK模型在留出验证中区分肿瘤与良性组织的整体准确率最佳（AUC约0.95；准确率约0.89）。相比之下，单个放大倍率的MUSK模型表现明显较差（AUC约0.90至0.92；准确率约0.82至0.85）。使用CONCH和HOPTIMUS训练的模型表现出类似趋势，多尺度模型达到AUC约0.94和准确率约0.89，相较于单一放大倍率模型（AUC约0.91，准确率0.81-0.84）有所改善。
结论：整合基础模型在多种尺度和放大倍率上的表征可实现数字病理中RCC肿瘤的精确定位。这些发现将在更大的RCC队列中进行验证，并评估其对治疗选择的影响。
查看英文原文 English abstract
Background: Renal cell carcinoma (RCC) diagnosis relies on accurate tumor localization on pathology whole slide images, but suffers from substantial inter-observer variability, which can impact treatment selection. Recent advances in digital pathology AI tools have resulted in foundation models (FMs) that are trained on diverse H&E images to learn patterns of tissue morphology. However, the performance of these models has been evaluated at a single scale, whereas clinical pathology evaluation leverages multiple scales and magnifications. Our objective was to develop and validate multi-scale pathology FM signatures for tumor localization in renal cancers.
Methods: H&E-stained whole slide images of RCC were curated from a public repository. For each case, tissue regions were sampled at three histologic scales (5×, 10×, and 20× magnification). Using the pathology FM MUSK, quantitative representations of tissue morphology were generated from each region and at each scale, and a supervised classifier trained on available pathologist annotations was used to discriminate tumor from non-tumor tissue. A multi-scale signature was constructed by integrating feature representations across all three magnifications for each region. We confirmed our findings using additional FMs (CONCH, HOPTIMUS) and by comparing against models trained at individual magnifications.
Results: A total of 118 slides (2 institutions) were included, split by patient into training (n = 94) and hold-out validation (n = 24) sets. Slides were sampled to yield 300,000+ labeled tiles (~112,000 tumor and ~222,000 non-tumor). The multi-scale MUSK model yielded the best overall accuracy for discriminating tumor versus benign tissue in hold-out validation (AUC ~0.95; accuracy ~0.89). By comparison, MUSK models at individual magnifications performed markedly worse (AUC ~0.90 to 0.92; accuracy ~0.82 to 0.85). Models trained using CONCH and HOPTIMUS demonstrated similar trends, with multi-scale models achieving AUC ~0.94 and accuracy ~0.89 that were improved relative to single magnification models (AUC ~0.91, accuracy 0.81-0.84).
Conclusions: Integrating foundation model representations across multiple scales and magnifications yields accurate tumor localization of RCC on digital pathology. These findings will be validated in larger RCC cohorts and evaluated for impact on treatment selection.
利益披露 Disclosure
S. Kapadia, None..
B. Flannery, None..
S. Viswanath, None.